Fast.ai notes

Auto-Regression.py

#Lib versions

#Python Version: 3.6.7
#Pandas Version: 0.24.2
#Numpy Version: 1.16.2
#FastAI Version: 1.0.50.post1
#PyTorch Version: 1.0.1.post2

#GPU Acceleration
#GPU: NVDIA K80
#Cuda V10.0.130
##############################
column_names = df.columns.values.tolist()
# (1) Separate  columns into continous data columns and category data columns 
cont_names = []
cat_names = []
cont_names = column_names[1:] # store everything except the first column "Date"

# (2) Converted "Date" column to categories, as suggested by the Docs
add_datepart(df, 'Date')
column_names = df.columns.values.tolist()
cat_names = column_names[6:len(column_names)-1] # Everything except the first seven and the last element

# inspect all meta data 
print("Nr. of all columns: " + str(len(column_names)))
print("Nr. of categorial data columns: " + str(len(cat_names)))
print("Nr. of continous data columns: " + str(len(cont_names)))
delta = len(column_names) - (len(cat_names) + len(cont_names))
print("Difference between all columns and all cat / cont columns: " + str(delta))

## Output 
#Nr. of all columns: 19
#Nr. of categorial data columns: 12
#Nr. of continous data columns: 7
#Difference between all columns and all cat / cont columns: 0

# (3) Added data processors
procs = [FillMissing, Categorify, Normalize]

# (4) Split data by index
valid_idx = range(len(df)-1000, len(df))

# (5)  a test data set as TabularList 
test = TabularList.from_df(df.iloc[:-2000].copy(), 
                           #path=path, 
                           cat_names=cat_names, 
                           cont_names=cont_names)
# (5) Created a databunch 
data = (TabularList.from_df(df,
                            #path=path, 
                            cat_names=cat_names, 
                            cont_names=cont_names, 
                            procs=procs)
                           .split_by_idx(valid_idx=valid_idx)
                           .label_from_df(cols=dep_var)
                           .add_test(test, label=0)
                           .databunch())

data.show_batch(rows=3)# looks okay, although I'm not sure wjhy the dependent variable is still there.

# (7) Created a tabular learne
learn = tabular_learner(data, layers=[200,100], metrics=accuracy_long)
# accuracy_long fixes a type casting error. I believe the dependent variable must be type long to avoid this fix. TODO for later

epochs = 8
lr = 1e-3
learn.fit(epochs)
## Output 
Total time: 00:09
epoch	train_loss	valid_loss	accuracy_long	time
0	3623.272461	536329.562500	0.000000	00:00
1	2522.631348	166452.062500	0.000000	00:00
2	1433.353516	15016.925781	0.000000	00:01
3	679.439087	12283.965820	0.000000	00:01
4	282.636536	20604.240234	0.000000	00:01
5	136.765137	9877.943359	0.000000	00:01
6	76.523834	6982.699219	0.000000	00:01
7	56.457199	17809.328125	0.000000	00:01